It is 11:40 at night and the seller has sixty-one unread messages in Telegram. Fourteen sit in the shop's group, nine hang under a channel post, and the rest are private chats. She opens the private ones first, because those feel like real buyers. By morning the group has answered itself: two of those answers are wrong, and one of them is about the refund policy.
Nobody warns you about that part. Telegram doesn't hand you an inbox. It hands you three of them, each with its own habits, and your customer has no idea which one they are standing in.
A telegram bot for customer support usually gets sold as a way to reply faster. Speed is the easy half. The half that decides whether any of this works is whether those three surfaces start behaving like one support desk with one memory.
Telegram support breaks in public
On Instagram, a question arrives in a DM and dies in a DM. Whatever you answer, one person reads it. Telegram is louder.
Someone asks "does the blue one come back in stock?" under a channel post, and every subscriber who scrolls past sees the question and the silence under it. Then the same person asks in the group. Then they message you privately, assuming the first two went nowhere. One question, three copies. Answer them slightly differently, which tired humans always do, and you have published a contradiction in a place where screenshots live forever.
Replying faster doesn't fix that. One source of truth answering all three, in the same words, every time, does.
Three surfaces, three jobs
Before automating anything, decide what each surface is for. Most of the support chaos I see comes from sellers treating all three as one undifferentiated inbox.
- Private chats. Where money moves. Order questions, sizes, payment problems, complaints. Anything involving a name, an address, or a receipt belongs here and nowhere else.
- Groups. Half community, half support queue. The valuable thing about a group is that one good answer gets read by two hundred people who never asked. The dangerous thing is that a wrong answer from a random member gets read by the same two hundred.
- Channel comments. Your shop window. Questions here are short and pre-purchase: price, availability, delivery time. They are also the most public, which means they should be the fastest.
The routing rule fits on a sticky note: public question, public answer; personal detail, private chat. A bot can enforce that at 3 a.m. You cannot.
Build the knowledge base first, the buttons later
The instinct is to start with a menu. Press 1 for orders, press 2 for shipping. Menus feel controlled, and they collapse the moment somebody types "hi, the black one you posted Tuesday, do you have 42?" — which is how real people write.
Start from the other end. Write down what your business actually knows, then let the bot answer from that in plain language. A useful knowledge base for an online seller is smaller than you expect:
- Products with sizes, materials and colours, plus what is genuinely out of stock right now
- Prices, payment methods, and whether you accept instalments
- Shipping cost and delivery time per region, and your cut-off hour for same-day dispatch
- The return window, who pays return shipping, and what voids it
- Warranty and after-sales, in the exact wording you want repeated
- How a customer checks the status of an order they already placed
Five questions cover most of the volume in a seller's inbox: price, availability, delivery time, size or fit, and some version of "is this the real thing?". Handle those five properly and the bot has earned its keep before you build anything clever.
One more requirement: the answers have to sound like you. A support bot writing like a bank leaflet, attached to a shop that writes like a friend, reads as fake — and customers spot it faster than you would like. The same principle runs through our guide to Telegram smart automation, and it applies to support just as much as to marketing.
The workflow that holds up
Here is the shape I would give a Telegram support setup for a small store, in order.
- Catch messages on all three surfaces. Private chats, group mentions and channel comments feed one queue. If your bot only watches private chats, you have automated the easiest third of the problem.
- Recognise who is asking. A returning buyer with an open order gets a different first sentence than a stranger asking about price. If your bot remembers nothing between messages, every conversation restarts from zero and customers repeat themselves.
- Answer from the knowledge base, in your voice. Not a canned template with the customer's name pasted in. An actual answer to the actual question, including the awkward ones like "why is it more expensive than the other shop?".
- Do something, don't only reply. Check an order status, register a complaint, add the person to a restock list, send the size chart. Support that only produces sentences pushes the work back onto you.
- Pull personal matters into private. A public "I've sent you a private message about your order" is polite, protects the customer's details, and clears the public thread.
- Hand off cleanly. When a human takes over, they should see the whole conversation, not a notification saying someone needs help. Handoffs that start with "hello, how can I help?" after the customer has already explained twice are worse than no bot at all.
- Log every question it could not answer. That list is your roadmap. Read it weekly, add the answers, and the bot's coverage climbs on its own.
Let it say "I don't know"
The fastest way to lose trust is a bot that guesses confidently. If a customer asks whether a jacket runs small and the bot invents an answer, you don't get a support ticket — you get a return, a refund and a bad review, in that order.
Decide up front what it must never handle alone:
- Refunds, complaints, and anything written in capital letters
- Money that has already been paid, especially failed or duplicated payments
- Wholesale, bulk pricing and custom orders
- The same question asked twice after the bot has already answered it — that repetition is the customer telling you the answer was wrong
- Anything with a legal, medical or safety claim attached to it
An honest "let me get a colleague to confirm that, one moment" costs you nothing. A confident wrong answer costs you the order.
What to watch in the first month
Numbers you can collect without a dashboard project:
- First response time, split by surface. Channel comments should be quickest; that is the surface strangers judge you by.
- Share of conversations closed without a human. Anything above half in month one is a good start. Chasing 100% is how bots get annoying.
- Handoff rate and handoff reasons. The reasons matter more than the rate.
- The unanswered-question log. Your top ten gaps, ranked by how often people hit them.
- Replies after 10 p.m. Night volume is real, and it is the volume you were losing.
Speed deserves a special mention because sellers routinely underestimate how brutal the clock is. We put real numbers against it in our comparison of AI and human reply speed, and the follow-up on how long DM customers actually wait is worth reading before you promise yourself you'll answer everything by hand.
The mistakes I keep seeing
- A menu tree pretending to be support. Six buttons, none of which is the question the customer has.
- Automating the group and ignoring channel comments. Comments are where undecided buyers stand.
- No handoff. The bot answers, the customer wants a person, and the conversation ends in a wall.
- A knowledge base written once and never touched. Prices change. Stock changes. A bot quoting last season's price is worse than silence.
- Hiding that it is a bot. You don't need a disclaimer in every message, but pretending to be a person breaks badly when the customer figures it out.
Start smaller than you want to
Pick your fifteen most common questions. Write real answers. Turn the bot on for channel comments only, watch it for a week, then add private chats, then the group. You will find gaps, and finding them one surface at a time is far less painful than finding them everywhere at once.
If you would rather not wire this together yourself, that is roughly what we built Vardast to do: connect your Telegram alongside Instagram, WhatsApp and your website, answer customers from your own product knowledge in your own voice, and pass the conversation to you the moment it needs a human. Same memory, every channel.
The seller with sixty-one unread messages didn't need to type faster. She needed the answers to exist in one place, and something willing to repeat them at midnight.